AI suggests. Editorial judgment is the discipline of ignoring most of what it suggests. The line between AI-assisted work and AI-generic work doesn’t run through the prompt, the model, or the knowledge base. It runs through what you choose not to keep.
This is the cluster closer for the AI & Workflows pieces on Freymwork. The earlier pieces drew lines around what AI is good at, when to put it in a workflow, how to recover when the plateau hits, and how to build a knowledge base that compounds. This one is about the move that protects all four upstream decisions from quietly failing. Without the discipline below, every other piece of the stack starts producing average output without anyone noticing.
Why “technically right” isn’t enough
The hardest AI suggestion to ignore is the one that’s technically right and wrong for the piece.
A more readable sentence. A clearer transition. A more accessible explanation. A more balanced framing. Each one of these is a defensible suggestion in isolation. Each one is also, often, the move that flattens the voice. Accept enough of them in a row and the prose stops sounding like anyone in particular. The model isn’t doing anything wrong. The model is producing the average defensible suggestion, which is what the average reader of the average piece would prefer. That’s not who you’re writing for.
This is the move that separates AI-assisted writing from AI-generic writing. The AI-assisted writer takes a suggestion that’s technically right and asks whether it’s right for this piece, this reader, this position. The AI-generic writer takes a suggestion that’s technically right and accepts it. After three months, only one of them still has a voice.
The distinction matters because it isn’t a question of competence. The generic version reads competently. The piece-specific version reads like work. The reader can tell the difference when the archive is compared at quarter-end, even if they couldn’t articulate it on any individual piece.
John McPhee’s essay on revision makes a related argument about choice and ignoring. McPhee describes circling a word he isn’t sure of and coming back to it days later, after the rest of the piece has settled, to decide whether the word belongs. The discipline is the deliberateness of the choice, not the speed of accepting the first plausible option. With AI, the temptation to accept the first plausible option is structural – the suggestion is right there, fully formed, and rejecting it requires more friction than accepting it. The discipline is rebuilding that friction back into the choice.
Three categories to default to ignoring
Three patterns of AI suggestion are wrong by default. Not always wrong. Always-default-wrong-until-proven-otherwise.
Smoothing of distinctive phrasing. A short, unconventional sentence becomes a slightly longer, more standard sentence. A phrasing you chose because it was specific to you becomes a phrasing that’s clearer to a stranger. The model is optimising for readability across the broadest possible audience. The distinctive phrasing was the part that signalled who wrote the piece. The smoother version reads more easily. It also reads more anonymously.
Generalisations that round off the edges. A specific claim about your industry, your reader, or your experience becomes a more general claim. “Most solo founders building solo content systems hit the plateau between month three and month six” becomes “Many creators experience challenges with content systems over time.” The first is a position. The second is a hedge. The model rounds off because the broader claim is harder to argue with. The position is what the piece was actually about.
Safe alternatives that displace the risky-but-right choice. A pointed observation becomes a neutral observation. A strong opinion becomes a balanced overview. The piece’s strongest moment becomes the piece’s most acceptable moment. The model defaults to safety because safety is the lowest-risk response across the population of users. The pointed version is what made the piece worth reading.
All three of these patterns share a property: they preserve what’s literally true while removing what was distinctive. The output passes any individual check – grammar, clarity, balance – while failing the only check that actually matters, which is whether the piece sounds like someone in particular wrote it.
Why the model defaults this way
The pattern isn’t accidental. The 2023 Sharma et al. paper on sycophancy in language models documented a measurable bias toward producing output the user is likely to agree with, even when the user’s position is wrong or the alternative is more accurate. The pattern compounds with editorial-suggestion tasks because the model is doubling down on a safety signal: it wants its suggestions accepted, so it produces suggestions that are easiest to accept.
The implication for solo work isn’t to distrust the model. It’s to recognise the bias as a structural property and adjust the acceptance threshold to compensate. A suggestion that’s easy to accept is not the same as a suggestion that’s right. The discipline is treating the ease as suspicious rather than confirming.
The earlier piece on the four jobs AI shouldn’t do drew the line at the category level. This piece draws it at the sentence level. Editing your voice was named as one of the four; the suggestions above are the texture of how that editing happens, sentence by sentence, when you let it.
The 30 percent rule
A workable threshold for editorial work with AI: accept 30 percent of what the model suggests at most.
The number is approximate, not absolute. The principle is that the acceptance rate matters as much as the quality of any individual acceptance. A high acceptance rate – 60, 70, 80 percent – produces an archive that drifts toward the average of all the small flattening choices. A lower rate – 20 to 30 percent – keeps the distinctive parts of the original draft intact while letting AI fix the things that genuinely needed fixing.
The 30 percent is harder than it sounds. Each individual suggestion looks defensible. Rejecting most of them feels stubborn or anti-AI. It isn’t either. It’s editorial judgment applied at the right ratio for the work to remain yours.
A practical test, when you’re not sure: read the next sentence in the AI-edited version and the original version aloud. The one that sounds like a sentence you would have written is the one to keep. The one that sounds like a sentence a competent editor might have produced is the one to reject. Both versions are defensible. Only one is yours.
This connects to the knowledge base discipline at a different scale. The knowledge base shapes what the model produces. The editorial threshold shapes what survives the model’s pass. Both are filters. Both compound.
Building the discipline
The discipline is not natural. The output is always there, always defensible, always ready to accept. Three small practices make the discipline easier to hold:
The first is structural delay. Don’t accept AI suggestions in real time. Let the suggestions sit for 24 hours and re-read both versions before deciding. The time gap surfaces which version actually sounded better. Real-time editing accepts too much because the immediate comparison favours the more polished prose. The 24-hour comparison favours the version that holds up after the polish wears off.
The second is reading aloud. Suggestions that look fine on the page often sound wrong when spoken. The voice the AI smooths into existence is a voice no one actually has. Reading aloud surfaces the gap. It also catches the suggestions that are technically right and rhythmically wrong – sentences that scan correctly but lose the cadence that made the original work.
The third is a quarterly audit. Pull ten pieces you’ve published in the last quarter and read them as if you were someone else encountering the archive for the first time. Ask: does this read like a body of work by one writer, or does it read like a series of competent pieces that could have been written by anyone? If the second answer is closer, the acceptance rate has been too high. Re-anchor to a lower threshold for the next quarter.
These three practices are the maintenance for the discipline. The discipline itself is the willingness to reject suggestions that look right because they aren’t right for the piece you’re trying to publish. The piece on editing your strategy, not your articles covers the strategy-layer version of this same loop – the editorial discipline that protects the archive’s direction, not just any individual piece’s voice.
When ignoring is wrong
The discipline of ignoring works because it’s a default, not a rule. Two situations override it.
The first is when the AI catches a factual error. A claim is wrong. A citation is broken. A statistic doesn’t match the source. The discipline doesn’t apply to corrections of fact. Take the correction.
The second is when the AI catches a structural problem the writer is too close to see. A confusing transition. A buried lead. A section that doesn’t earn its place. These are not voice issues – they’re craft issues. The discipline applies to voice and judgment, not to the technical execution of the piece. If AI flags that an argument loses its way, the appropriate response is to fix the argument, not to defend the original.
The way to tell the difference between the two situations and the default-ignore cases: corrections of fact and structural problems will keep producing the same flag across multiple readings. Voice and judgment suggestions will only produce the flag once. The persistence of the suggestion is the signal that it might be one of the two override cases.
What the cluster ends with
The five pieces in this cluster cover the AI work-frame at different layers. What AI is actually good at draws the category line. The five-question pre-flight draws the workflow line. The plateau-crossing patterns draw the maintenance line. The knowledge base draws the compounding line. This piece draws the line at the sentence – the final filter that determines whether the rest of the stack produced a body of work or a body of competent output.
The line at the sentence is the hardest to hold because every individual rejection feels small. The compounding is real. A year of rejecting AI’s smoother alternatives produces an archive that sounds like a person. A year of accepting them produces an archive that sounds like the average of the AI-assisted writing on the internet, which is not what your reader came for.
The rest of the AI & Workflows archive lives at the AI & Workflows category.
The work is the ignoring
AI’s job is to produce options. Yours is to ignore most of them. That isn’t laziness. It’s editorial judgment applied at the sentence level, accumulated across a year, compounding into an archive that sounds like one writer rather than the average of all writers.
The piece that survives the AI-edit pass is the piece that holds. The work is the ignoring.
Build Your Content Machine. A free 5-part email course on building an AI content system that sounds like you – the discipline above is the filter it’s built around. Start the free course →
